Calling Variants - iffatAGheyas/bioinformatics-tutorial-wiki GitHub Wiki

Calling Variants

Pileup-based callers (bcftools mpileup + call)

1. Exit your InSilicoSeq virtual environment (if you’re still in it):

deactivate

2. Install bcftools and tabix (needed to call and index VCFs):

   sudo apt update
   sudo apt install -y bcftools tabix
   

3. Re-enter your project directory and (re)activate your ISS env:

   cd /mnt/c/Users/IAGhe/OneDrive/Documents/Bioinformatics
   source iss_env/bin/activate
   

4. Run the pileup → call pipeline:

   bcftools mpileup \
     -f ref_genome.fa \
     -Ou \
     aln.filtered.bam \
   | bcftools call \
     -mv \
     -Oz \
     -o raw_snps_indels.vcf.gz
   

5. Index the VCF (for fast querying):

   tabix -p vcf raw_snps_indels.vcf.gz
   

6. Take a peek at your variants:

   bcftools view raw_snps_indels.vcf.gz | head -n 30
   

7. (Optional) View in a nice column format:

   bcftools view raw_snps_indels.vcf.gz \
     | column -t \
     | less -S
   

All of the above is adapted from the InSilicoSeq hands-on demo.

---

Haplotype-aware callers

• GATK HaplotypeCaller

# Requires GATK installed (via Conda or you can download the jar)
gatk HaplotypeCaller \
  -R ref_genome.fa \
  -I aln.filtered.bam \
  -O gatk_variants.vcf.gz

• FreeBayes

# A lightweight, multiallelic haplotype caller
freebayes \
  -f ref_genome.fa \
  aln.filtered.bam \
  > freebayes_variants.vcf

(Commands to run these would follow a similar pattern: index the BAM, then invoke the caller with -R ref_genome.fa, -I aln.sorted.bam, and appropriate output flags.)

---

Long-read callers

• Medaka (for ONT data)

# After mapping ONT reads with minimap2
medaka_variant \
  -i ont_reads.fastq.gz \
  -d ref_genome.fa \
  -b ont_aln.bam \
  -o medaka_variants

• Clair3

# Clair3 Docker or conda install; specify platform 'ont' or 'hifi'
clair3.sh \
  --bam ont_aln.bam \
  --ref ref_genome.fa \
  --threads 8 \
  --platform ont \
  --output clair3_out

(These tools take long-read alignments (e.g. minimap2 → BAM), then call variants using neural-network models trained on long-read error profiles.)